Two AI personas in one image without merged faces

To keep two fixed AI personas apart in a single image, give each persona its own anchor: its own reference face, a clearly contrasting look, and a fixed spot in the frame. Without those anchors, two generated faces quickly start to resemble each other, or swap features. Below you'll read why that happens and how to place two recognizable identities side by side.

Why two faces drift toward each other

Models merge identities because they spread features across the whole scene instead of per person. This is called identity blending: the traits of two subjects bleed into each other, so you get two half-brothers instead of two different people.

It's a known, measured problem. The MultiHuman-Testbench (2025), the first large-scale benchmark for image generation with multiple humans, reports that current methods regularly show identity blending, generate the wrong number of humans, or fail to compose the scene according to the text.

Two things make it harder. Two humans already look more alike than a human and a dog, so there is less difference to hold on to. And if you describe them almost identically ("two women in their twenties"), you give the model nothing to separate them with.

Give each persona its own anchor

The most reliable step is to give each persona its own contrasting look. The bigger the difference, the less the model is tempted to average the faces together.

• Choose different hair: color, length, and style.
• Vary build, skin tone, and age where it fits your personas.
• Give each a distinct clothing style and one recognizable accessory.
• Fix their positions: one on the left, one on the right.

Modern premium models can hold more identities at once than a year ago. Google's Nano Banana Pro (Gemini 3 Pro Image), one of the models in the photo generator, keeps a consistent resemblance for up to five people according to Google, and blends up to fourteen reference images into a single image. So two personas side by side is well within reach, as long as you give each face its own reference instead of one shared description.

Describe them separately, not as one group

Describe each persona as its own block, not as a pair. State who stands where first, then the traits per person.

A workable structure:

1. Position: "a woman on the left, a man on the right".
2. One sentence per person with hair, outfit, and accessory.
3. The scene and light last, shared by both.

Avoid a single loose list where traits float around ("blond hair, red jacket, glasses, beard"). The model then doesn't know which trait belongs to whom, and you end up with attribute swapping: the glasses jump to the wrong face. Tie every trait explicitly to one person.

If it still fails: generate separately and combine

If a single render doesn't work, generate the personas separately and bring them together. That is often faster and cheaper than re-rolling until both faces happen to line up.

• Generate persona A in the scene and leave room for persona B.
• Add persona B with inpainting in the photo editor, or composite two separate images.
• Match light, shadow, and scale so the two don't look like separate cut-outs.

Because you pay per render, editing one region is usually cheaper than generating the whole image again.

Check the identity on every render

Put your two reference faces next to the result and check that both still match. Small drifts stack up across a series, especially if you later turn the images into video. Keep one strong reference per persona and reuse it, so your duo stays recognizable across multiple posts.

Frequently asked questions

Why do my two AI characters keep looking alike?

Because the model spreads traits across the whole scene, and two humans already look similar. Give each persona a clearly different look and its own reference face, and the model will keep them apart more easily.

Can one model hold two identities at once?

Yes, to a point. Google's Nano Banana Pro keeps a consistent resemblance for up to five people. So two personas works, as long as you supply each face separately instead of through one shared description.

What do I do if the faces still merge?

Generate the personas separately and add the second one with inpainting, or composite two images. That is usually faster and cheaper than trying the whole image again.

Two recognizable personas side by side comes down to making a difference and anchoring each face separately. Give both their own reference, describe them separately, and fix things where the model hesitates. Create an account and place your first duo side by side.